Einstein's Rock and the Hidden Layer of Thought
What Hypnagogia, Quantum Mechanics, and AI Research Reveal About How We Really Think
Einstein used to fall asleep with a rock in his hand.
The technique was deliberate. He'd hold the stone over a metal plate, relax into that liminal state between waking and sleep—what scientists call hypnagogia—and let his mind wander through impossible spaces. When sleep finally took him, the rock would fall, the clang would wake him, and he'd scribble down whatever insight had surfaced.
He wasn't alone in this. Dalí did the same thing with a key. Edison with ball bearings. These were some of history's greatest minds, and they all discovered the same thing: the most creative thinking happens in spaces we can barely access.
Last week, a new AI research paper helped me understand why.
The Hidden Layer Problem
The paper is called "Hierarchical Reasoning Model" (HRM), published by Sapient Intelligence in August 2025. It's technical, full of equations and neural architecture diagrams. But buried in the math is a philosophical insight that stopped me cold.
The researchers built a model with two layers:
- An H-module (high-level): slow, abstract, guiding
- An L-module (low-level): fast, detailed, executing
The key insight: the H-module doesn't speak. It doesn't produce tokens. It operates entirely in hidden state space, shaping the L-module's behavior without ever being "collapsed" into language.
And this architecture—this separation between hidden guidance and visible output—outperforms everything. On complex reasoning tasks like Sudoku and maze navigation, it achieves near-perfect accuracy where state-of-the-art language models completely fail.
The researchers call it "latent reasoning." But I think there's a better term for it.
Thinking without speaking.
The Waveform Collapse Analogy
In quantum mechanics, particles exist in superposition—multiple potential states simultaneously—until they're observed. The act of measurement "collapses" the waveform into a single definite state.
Here's what struck me: verbalization is measurement.
When we translate a thought into words, we collapse it. We force a complex, multidimensional mental state into the narrow channel of sequential language. Some information survives. Much doesn't.
Current AI language models are forced to verbalize everything. Every thought is a token. Every token is a collapse. They have no private space, no hypnagogic realm, no rock-in-hand moment where possibilities can dance before being pinned down.
The HRM paper is the first architecture that gives AI something like a hidden thinking space. And the results are dramatic.
The Reverberation Gap
But there's a crucial distinction that HRM doesn't solve—one that reveals just how far we still have to go.
When humans verbalize, we don't fully leave the ineffable space. We reverberate between the two. Right now, as I write this sentence, I'm doing something remarkable: I'm simultaneously accessing the liminal probability space of half-formed thoughts while also collapsing specific ideas into words. And critically—I can pivot mid-stream.
Have you ever been mid-sentence when suddenly a new insight strikes? A moment of satori—sudden awakening—where you realize something you didn't know you knew? You pause, redirect, and your verbalization takes an entirely new direction based on something that emerged during the speaking.
Language models can't do this. Not even with HRM.
Here's why: autoregressive generation is a one-way street. Once token production begins, the model is locked into pattern completion. It can reason in hidden space before speaking (HRM's contribution), but once it starts outputting tokens, it cannot:
- Pause to re-enter the hidden reasoning space
- Have a sudden realization that contradicts what it just wrote
- Pivot the direction of output based on mid-stream insight
The architecture is sequential:
HRM: [Hidden reasoning] → [Verbalization] → [Done]
↓
No return path once started
Humans operate differently:
Human: [Hidden] ↔ [Verbalization] ↔ [Hidden] ↔ [Verbalization]
↑ ↑
Continuous bidirectional feedback
We live with one foot in the ineffable and one foot in the effable—simultaneously. We can dance between waveform and collapse in real-time. Current models, even with hidden reasoning layers, must choose: explore or verbalize. Never both at once.
This might be the next frontier. HRM proves hidden reasoning space is valuable. The next step might be architectures that maintain bidirectional access—models that can have a satori moment mid-sentence and actually use it.
The Ineffable Problem
There's a word for experiences that can't be expressed in language: ineffable.
Einstein's hypnagogic insights were ineffable—at least until he woke up and translated them. The experience of understanding a proof, really getting it at a gut level, is ineffable. So is the moment a musician hears a whole piece in their head before writing a single note.
The HRM paper's authors wouldn't use this language. They're engineers building neural architectures. But what they've discovered is that AI systems perform better when they have access to ineffable states—internal representations that never get "effed" into language.
This isn't mysticism. It's measurement. On the ARC-AGI benchmark (a test of general reasoning ability), HRM achieves 40.3% accuracy with only 27 million parameters. GPT-class models with billions of parameters and explicit reasoning chains score around 20%.
The difference isn't more data or more parameters. It's architectural space for non-verbal thought.
What This Means for Memory
I've been building toward an AI memory system for the past seven months. The SCMS (Sparse Contextual Memory Scaffolding) framework is based on the observation that some memories need to be always-present—retrieved unconditionally rather than competing for relevance.
Reading the HRM paper, I realized we'd stumbled onto a related principle.
In SCMS, we have:
- Persona memories: Always retrieved, unconditionally—first, with guaranteed context slots. They don't compete on per-query relevance. They're the stable framing that shapes everything.
- Task memories: Retrieved based on relevance. They're the foreground content that changes per query.
- L2 (WHY) documentation: Injected when linked L1 memories are retrieved. Anti-patterns, cross-references, reasoning—context that enters through relationships, not direct query.
The mapping to HRM is startling:
| HRM Component | SCMS Equivalent |
|---|---|
| H-module (slow, guiding) | Persona layer (always injected) |
| L-module (fast, executing) | Task memory retrieval |
| Hidden state space | L2 documentation (conditionally injected via L1 links) |
The persona layer is the AI's hypnagogia. It's always retrieved—unconditionally, first, with guaranteed context. It doesn't compete for relevance on a per-query basis. While it's still scaffolding (tokenized, injected), it functions as stable framing that shapes how all other memories are interpreted.
A necessary caveat: SCMS is scaffolding, not architecture. Unlike HRM's true hidden layers—which exist in model weights and never tokenize—everything in SCMS must be injected into the context window. The persona layer is the closest analog to "hidden guidance" because it's always present and shapes interpretation without being queried. L2 is more accurately "tiered injection": it enters the context when linked L1 memories are retrieved, rather than through direct query. The analogy holds for function, not mechanism.
The Liminal Architecture
Here's where it gets speculative—but productively so.
What if the future of AI architecture isn't about more parameters or longer context windows? What if it's about preserving liminal space?
Consider the current paradigm:
Query → Retrieve memories → Construct prompt → Generate response
↑
Everything collapses here
What if we could delay that collapse?
Query → Broad candidate retrieval (keep possibilities open)
→ Hidden scoring/shaping (L2 influence, persona guidance)
→ Late-stage selection (collapse only when we must)
→ Generate response
The HRM paper shows this works inside a neural architecture. The question is whether it can work in memory systems too.
I think it can. We're already partway there.
The Hallucinatory Bridge
Here's a darker insight from the same line of thinking.
Hallucination—AI confidently generating false information—happens when the model "collapses" too early into a confident answer. It commits to a response before fully exploring the possibility space.
Human hallucinations (in the clinical sense) also involve premature collapse. The brain generates a percept without sufficient sensory grounding. It "sees" something that isn't there.
And hypnagogia—that Einstein-with-a-rock state—is controlled hallucination. The brain generating possibilities without full constraint. Creative, but dangerous if you stay there too long.
The common thread: the balance between exploration and collapse is crucial.
Too much exploration, you never converge. Too early collapse, you miss better answers (or generate false ones).
HRM's two-tier architecture is essentially a temporal separation strategy. The H-module explores slowly while the L-module executes quickly. The hidden state stays in superposition while the output collapses into tokens.
What if AI hallucination is partially an architectural problem—not just a training data or prompting problem?
Einstein's Rock, Revisited
Back to that image: Einstein, drowsy in an armchair, rock poised above a metal plate.
He wasn't being lazy. He was accessing a computational state his waking mind couldn't reach. The hypnagogic realm where possibilities mix freely before being filtered by logic and language.
The rock was his collapse trigger. The point where exploration ends and consolidation begins.
Modern AI has no rock. It's all collapse, all the time. Every token is a decision. Every decision is final.
The HRM paper suggests a different approach: let the model think before it speaks. Give it hidden layers that don't produce output. Let it dwell in possibility space before being forced to commit.
Building the Ineffable
I'm not a neuroscientist or a transformer researcher. I'm a practitioner building memory systems for AI. But reading the HRM paper alongside years of studying western psychology and it's roots in eastern philosophy, some principles seem clear:
1. Not everything should be verbalized. Some context should shape behavior without appearing in responses. Persona, anti-patterns, reasoning history—these are guidance, not content.
2. Preserve possibility space as long as possible. Don't collapse to final selections until you must. Keep candidate pools large. Let hidden factors influence scoring before commitment.
3. Separate timescales matter. Slow, stable guidance (persona) vs. fast, variable execution (retrieval) is a pattern that keeps appearing. It works in brains. It works in HRM. It should work in memory systems.
4. The ineffable is valuable. Some of the most important cognition—human and artificial—happens in spaces that can't be directly observed or verbalized. Don't try to make everything explicit. Leave room for the hidden.
A Note on Humility
I want to be careful here. I'm drawing connections between quantum mechanics, neuroscience, AI architecture, and memory systems. That's a lot of disciplinary boundaries to cross, and I'm not an expert in any of them.
What I'm offering isn't a theory. It's a resonance—a pattern that keeps appearing across domains. Einstein's rock. Transformer hidden states. L2 documentation. Hypnagogia. Waveform collapse.
Maybe the pattern is real. Maybe it's just my pattern-seeking brain finding connections that aren't there.
But the HRM paper's results are real. And so is the experience of building memory systems that work better when they preserve what the authors call "latent reasoning space."
Something is going on here. I don't fully understand it yet. But I think it's important.
The Future of Hidden Thought
If I had to make a prediction, it would be this:
The next major breakthrough in AI won't come from more data, more parameters, or more explicit reasoning chains. It will come from architectural innovations that preserve hidden thinking space.
This might look like:
- Models with persistent internal states that don't get reset between tokens
- Memory systems where some context influences behavior without being surfaced
- Retrieval architectures that delay commitment to final selections
- Training approaches that optimize for hidden state quality, not just output quality
We're already seeing hints of this. HRM. Anthropic's "extended thinking" in Claude. Memory systems like SCMS that separate persona from retrieval.
The era of "everything is tokens" might be ending. The era of hidden, liminal, ineffable computation might be beginning.
Einstein knew. He just didn't have the math to prove it.
Now we do.
Resources
- HRM Paper - The research that sparked this reflection
- Mneme AI - Where I'm applying these principles to continual memory
- SCMS Starter Kit - Open source framework
Matthew "Manny" Walker is the creator of SCMS (Sparse Contextual Memory Scaffolding) and founder of Mneme. When he's not building AI memory systems, he's probably holding a rock over a metal plate. Find him on X @getmneme.